FYI: Hardware for next generation artificial neural networks (ANNs)

Today’s ANNs use a lot of electrical power especially for training. This is creating a big demand on the national grid and acts as a downside for the expanded use of AI. The brain uses just a few watts, perhaps 20W, and is five orders of magnitude more power efficient than GPUs. The brain features local analog processing and spikes for long range communication. Can we match that?

A very promising approach is to use light pulses conveyed via optical waveguides as a replacement for electrical conductors within and between integrated circuits. This is faster and uses less power. We can build upon the extensive experience with CMOS fabrication with silicon on insulator (SOI) for hybrid photonic devices that combine electronic and optical components in the same chip. Silicon is transparent for wavelengths above 1.1 microns and has a very large refractive index (3.5) compared to silicon dioxide (1.4). Silicon's indirect bandgap makes it unsuitable for LEDs and photo diodes. One way to get around that is use InAs/GaAs grown on silicon substrates, however, a more promising approach uses germanium on silicon for infrared detectors, e.g. telecom's wavelengths such as 1.3 microns and 1.5 microns. Strained germanium on silicon can be used for emitters, but a more promising approach is to use germanium-tin alloys (GeSn).

Another potential approach is to use quantum dots which trap electrons and have discrete excitation bands similar to individual atoms. Silicon quantum dots (SiQD) with a size of 1 to 10 nanometres can be used for emitting and detecting light in a wide range of wavelengths from the near infrared to red/orange, but not so far for wavelengths where bulk silicon is transparent (wavelengths greater than 1.1 microns).

It might seem reasonable to look for ways to fabricate wave guides for red light where silicon is opaque. Silicon dioxide is highly transparent for visible light, but has the disadvantage of a much lower refractive index compared to bulk silicon. To avoid leaky waveguides, the cladding material needs to have a lower refractive index. Air is fine for this, but silicon is not. This makes it preferable to stick to infrared light where silicon is transparent. SiQDs can also be used for beam splitters and other optical devices. Silicon waveguides, just nanometres across, can efficiently transport infrared photons, even around tight bends, enabling a high density of communication paths.

For computing we need memory. Resistive random access memory (ReRAM) is based upon memristors, and can be integrated alongside CMOS electronic gates and optical waveguides. Memristors have a resistance that depends on the voltage and current previously applied to them. Memristors could be used to mimic synapses for neuromorphic computing, and can be fabricated using amorphous silicon doped with oxygen or nitrogen, or by using silicon dioxide layers on silicon hydride surfaces. Memristors are much smaller than flash memory cells and simpler to drive. Silicon foundries are moving towards 2 nanometre resolution using extreme ultraviolet (EUV) lithography, which should be sufficient for hybrid photonics.

Of course there are many challenges to address before this technology can become widespread. This includes surface passivation and physical encapsulation within a protective matrix, as well as connectivity to photonic chips compared to the ease of bonding wired connections. We are also likely to need new architectures for neural networks that feature continual learning through continual prediction, and locally generated weight adjustments layer by layer, replacing today’s gradient-based back propagation. Another motivation is that adversarial attacks suggest that today’s AI is different from human cognition, e.g. adding some carefully chosen noise to an image of a panda causes the model to misclassify it as a gibbon with a high degree of confidence, despite the noise being imperceptible to the human eye. Similar considerations apply to prompt injection jail breaks that evade the alignment training for LLMs.

Dave Raggett <dsr@w3.org>